todo-files/test_classif_mxff.R

In mlr-org/mlr: Machine Learning in R

test_that("classif_mxff", {
  requirePackagesOrSkip("mxnet", default.method = "load")
  # tests for FeedForward networks with only FullyConnected layers using mx.mlp
  parset.list.mxnet = list(
    list(hidden_node = c(10, 6), activation = c("sigmoid", "relu"),
      learning.rate = 0.2, array.layout = "rowmajor"),
    list(hidden_node = c(10, 6), activation = c("sigmoid", "relu"),
      learning.rate = 0.2, dropout = 0.5, array.layout = "rowmajor")
  )
  parset.list.mlr = list(
    list(layers = 2, num.layer1 = 10, num.layer2 = 6, act1 = "sigmoid",
      act2 = "relu", learning.rate = 0.2),
    list(layers = 2, num.layer1 = 10, num.layer2 = 6, act1 = "sigmoid",
      act2 = "relu", learning.rate = 0.2, dropout.global = FALSE, dropout.layer1 = 0.5,
      dropout.layer2 = 0.5)
  )

  # test for Convolution
  parset.list.conv = list(
    list(num.filter = 1, kernel = c(1, 1), pool.type = "max", learning.rate = 0.1)
  )
  parset.list.mlr.conv = list(
    list(layers = 1, conv.layer1 = TRUE, num.layer1 = 1, conv.data.shape = c(2, 2),
      conv.kernel1 = c(1, 1), pool.kernel1 = c(1, 1), pool.type1 = "max"
    )
  )

  # binaryclass tests
  old.predicts.list = list()
  old.probs.list = list()

  set.seed(getOption("mlr.debug.seed"))
  for (i in seq_along(parset.list.mxnet)) {
    x = data.matrix(binaryclass.train[, -ncol(binaryclass.train)])
    y = as.numeric(binaryclass.train[, ncol(binaryclass.train)]) - 1
    pars = c(parset.list.mxnet[[i]], list(data = x, label = y), out_node = 2)
    m = do.call(mxnet::mx.mlp, pars)
    probs = predict(m, data.matrix(binaryclass.test[, -ncol(binaryclass.test)]), array.layout = "rowmajor")
    p = apply(probs, 2, function(i) {
      w = which.max(i)
      return(ifelse(length(w > 0), w, NaN))
    })
    # avoid error when only one class is predicted
    p = factor(p, exclude = NaN)
    levels(p) = binaryclass.class.levs
    p2 = t(probs)
    colnames(p2) = binaryclass.class.levs
    old.predicts.list[[i]] = p
    # only save probs of positive class
    old.probs.list[[i]] = p2[, binaryclass.task$task.desc$positive]
  }

  set.seed(getOption("mlr.debug.seed"))
  testSimpleParsets("classif.mxff", binaryclass.df, binaryclass.target, binaryclass.train.inds,
    old.predicts.list, parset.list.mlr)
  set.seed(getOption("mlr.debug.seed"))
  testProbParsetsWithTol("classif.mxff", binaryclass.df, binaryclass.target, binaryclass.train.inds,
    old.probs.list, parset.list.mlr)

  # multiclass tests
  old.predicts.list = list()
  old.probs.list = list()

  set.seed(getOption("mlr.debug.seed"))
  for (i in seq_along(parset.list.mxnet)) {
    x = data.matrix(multiclass.train[, -ncol(multiclass.train)])
    y = as.numeric(multiclass.train[, ncol(multiclass.train)]) - 1
    levs = levels(multiclass.df[[multiclass.class.col]])
    pars = c(parset.list.mxnet[[i]], list(data = x, label = y), out_node = length(levs))
    m = do.call(mxnet::mx.mlp, pars)
    probs = predict(m, data.matrix(multiclass.test[, -ncol(multiclass.test)]), array.layout = "rowmajor")
    p = apply(probs, 2, function(i) {
      w = which.max(i)
      return(ifelse(length(w > 0), w, NaN))
    })
    # avoid error when only one class is predicted
    p = factor(p, exclude = NaN)
    levels(p) = levs
    p2 = t(probs)
    colnames(p2) = levs
    old.predicts.list[[i]] = p
    old.probs.list[[i]] = p2
  }

  set.seed(getOption("mlr.debug.seed"))
  testSimpleParsets("classif.mxff", multiclass.df, multiclass.target, multiclass.train.inds,
    old.predicts.list, parset.list.mlr)
  set.seed(getOption("mlr.debug.seed"))
  testProbParsetsWithTol("classif.mxff", multiclass.df, multiclass.target, multiclass.train.inds,
    old.probs.list, parset.list.mlr)

  # Convoution test
  conv.probs.list = list()
  set.seed(getOption("mlr.debug.seed"))
  x = data.matrix(multiclass.train[, -ncol(multiclass.train)])
  y = as.numeric(multiclass.train[, ncol(multiclass.train)]) - 1
  x = array(aperm(x), dim = c(2, 2, 1, nrow(x)))
  sym = mxnet::mx.symbol.Variable("data")
  sym = mxnet::mx.symbol.Convolution(sym, num.filter = parset.list.conv[[1]]$num.filter,
    kernel = parset.list.conv[[1]]$kernel)
  sym = mxnet::mx.symbol.Pooling(sym, kernel = parset.list.conv[[1]]$kernel,
    pool.type = parset.list.conv[[1]]$pool.type)
  sym = mxnet::mx.symbol.flatten(sym)
  levs = levels(multiclass.df[[multiclass.class.col]])
  sym = mxnet::mx.symbol.FullyConnected(sym, num_hidden = length(levs))
  out = mxnet::mx.symbol.SoftmaxOutput(sym)
  m = mxnet::mx.model.FeedForward.create(out, X = x, y = y, array.layout = "colmajor",
    learning.rate = parset.list.conv[[1]]$learning.rate)
  x.test = data.matrix(multiclass.test[, -ncol(multiclass.test)])
  x.test = array(aperm(x.test), dim = c(2, 2, 1, nrow(x.test)))
  probs = t(predict(m, x.test))
  colnames(probs) = levs
  conv.probs.list[[1]] = probs

  set.seed(getOption("mlr.debug.seed"))
  testProbParsetsWithTol("classif.mxff", multiclass.df, multiclass.target, multiclass.train.inds,
    conv.probs.list, parset.list.mlr.conv, tol = 1e-03)
})